Nvidia Finds $500B, Lovable Doubles, and DeepSeek Goes Premium

Nvidia recruited six of the world's biggest financial institutions to help create more than $500 billion of AI compute financing. Databricks raised $5 billion at a $190 billion valuation while growing revenue by more than 80%. Lovable doubled its valuation to $13.3 billion and announced that it is hiring heavily across machine learning, product, infrastructure, and security. DeepSeek then reminded everyone that stronger models do not stay cheap forever, pricing its new flagship up to 14 times above its faster alternative.

The common thread is not simply that AI is still attracting money. Capital is concentrating around the companies that own a genuine bottleneck: compute, enterprise data, software creation, model economics, security, or access to a strategically important market.

For hiring teams, that means the useful question is changing from "Do we need more AI engineers?" to "Which constraint in our AI system actually deserves another person?"

The Drop

1. Nvidia wants to turn AI compute into a $500 billion asset class

What happened: Nvidia signed memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to create compute-financing platforms targeting more than $500 billion of third-party capital. Nvidia said it could backstop up to $125 billion, or 25% of potential transactions. Big Tech's combined AI spending is now expected to exceed $730 billion this year. The individual commitments, financial terms, and deployment timetable have not yet been disclosed.

The risk is already visible. Later in the week, Nvidia was reported to have reduced its proposed initial guarantee for OpenAI's 10-gigawatt Ohio data-centre project from the previously discussed $250 billion to less than $120 billion after investors raised concerns about its exposure.

Why it matters for hiring: AI infrastructure is becoming part technology programme, part project finance, and part asset-management problem. The people deciding what to build now need to understand utilisation, power, residual GPU value, lease structures, customer concentration, and what happens if demand or chip performance changes faster than the debt matures.

Roles likely to benefit:

• AI infrastructure and data-centre programme leaders

• Capacity planning and GPU fleet engineering

• FinOps and AI unit-economics specialists

• Power procurement and grid specialists

• Infrastructure finance, risk, and hardware lifecycle teams

• Reliability engineers who can improve useful output per unit of compute

Hiring takeaway: A senior AI infrastructure interview should now include economics. Ask candidates how they would compare buying, leasing, reserving, or using cloud capacity, and what evidence would make them reduce a long-term commitment.

2. Databricks raises $5 billion as the enterprise data layer keeps winning

What happened: Databricks raised $5 billion at a $190 billion valuation, up from roughly $134 billion only six months earlier. It has passed a $7 billion annualised revenue run-rate, grew revenue by more than 80% year on year in the second quarter, and remained adjusted cash-flow positive across the last 12 months. Lakebase has exceeded a $100 million revenue run-rate, while the Lakehouse warehousing business has passed $1.5 billion.

Why it matters for hiring: The market is rewarding the layer that makes enterprise AI usable, governed, and connected to live company data. Models may change every few months. Data permissions, lineage, reliability, retrieval, and production workflows remain difficult.

Roles likely to benefit:

• Distributed systems and database engineers

• Data platform and streaming engineers

• AI gateway, inference, and agent-platform engineers

• Data governance, privacy, and lineage specialists

• Enterprise solution architects and forward-deployed engineers

• Developer experience and platform product leaders

Hiring takeaway: "Built an LLM feature" is no longer enough signal for a senior hire. Look for evidence that a candidate moved messy, permissioned, business-critical data into a reliable production workflow with measurable adoption.

3. Lovable doubles to $13.3 billion and gives us an unusually clear hiring map

What happened: Lovable raised $400 million in Series C funding at a $13.3 billion valuation, double its December valuation. Its annual recurring revenue has nearly tripled from $200 million and is tracking towards $600 million by the end of August. More than 60 million projects have been created since launch, with Lovable-built apps attracting over 900 million visits per month. Most importantly for this newsletter, the company plans to grow to roughly 450 people this year, hiring most heavily in machine learning, product, infrastructure, and security.

Why it matters for hiring: AI coding has not removed the need for engineers. It has moved demand towards people who can make software generation dependable at scale. When millions of users can create applications, the scarce work becomes architecture, platform reliability, secure defaults, payments, governance, observability, and turning prototypes into products that survive real traffic.

Roles likely to benefit:

• ML and agent-product engineers

• Platform, infrastructure, and site reliability engineers

• Application and cloud security engineers

• Product engineers who can own an outcome end to end

• Enterprise governance and integration specialists

• Developer experience engineers

Hiring takeaway: The strongest software candidates will increasingly be judged on what they can safely own, not how many lines they can manually produce. Add questions on generated-code review, dependency risk, observability, rollback, and long-term maintainability.

4. DeepSeek proves that better AI performance can carry a very large premium

What happened: DeepSeek released V4 Pro at $1.32 per million input tokens and $3.96 per million output tokens. That is about nine times the input price and 14 times the output price of V4 Flash. Artificial Analysis scored the reasoning version of V4 Pro at 53 on its Intelligence Index, compared with 40 for Flash. DeepSeek has also said it intends to at least double staffing across departments, including AI-agent and data-centre teams, while expanding private hiring for chip design.

Why it matters for hiring: The "models will keep getting better and cheaper" assumption is too simplistic. Providers can charge a large premium when better reasoning, coding, tool use, or reliability creates more business value. Teams therefore need to measure quality-adjusted cost, not token price alone.

Roles likely to benefit:

• Model evaluation and benchmarking engineers

• AI FinOps and inference optimisation specialists

• Model-routing and fallback-platform engineers

• Agent reliability and observability engineers

• Compiler, runtime, and chip-design engineers

Hiring takeaway: Give AI candidates two models with different quality, latency, and token prices. Ask them to design the evaluation and routing policy that decides which model handles each task. The best answer should optimise cost per successful outcome, not cost per token.

5. Apple is building a separate AI stack for China

What happened: Apple has reportedly trained a China-specific large language model with support from Alibaba. Apple Intelligence is expected to launch in China in the coming months following regulatory clearance, with Alibaba's Qwen also incorporated into the local product. If the proprietary model launches as described, Apple would become the first foreign company approved by Beijing to offer its own AI model in China.

Why it matters for hiring: Global AI products are becoming portfolios of regional systems. Regulation, model availability, local language performance, data rules, commercial partnerships, and geopolitical alignment can all change the technical stack by market.

Roles likely to benefit:

• Multilingual model evaluation and localisation

• AI policy, governance, and privacy engineering

• Regional data and safety teams

• Partner engineering and technical programme management

• Product leaders experienced in regulated markets

Hiring takeaway: If a product operates internationally, test whether candidates can design for regional model substitution, data residency, different safety requirements, and inconsistent feature availability without creating an unmaintainable fork for every country.

AI Tool of the Week

Screenloop

What it does: Screenloop combines an applicant tracking system with interview intelligence. Its AI notetaker records and transcribes interviews, detects question-and-answer segments, summarises evidence against predefined attributes, and drafts scorecards for review. It also supports interview clips, action items, interviewer coaching, and hiring analytics.

Who it is for: In-house talent teams that lose time chasing interview feedback or find that every interviewer interprets the same competency differently.

Seven-day pilot:

1. Pick one role with at least 10 interviews scheduled.

2. Agree five competencies and anchored scoring criteria before the first interview.

3. Use Screenloop to capture the conversation and draft each scorecard.

4. Require the interviewer to approve, edit, or reject every AI-generated assessment.

5. Hold one calibration session using anonymised evidence from three interviews.

Metrics to track:

• Median time from interview end to completed scorecard

• Percentage of scorecards completed within 24 hours

• Percentage of AI-drafted ratings changed by the interviewer

• Evidence coverage across the five competencies

• Agreement between interviewers at the debrief

• Candidate experience score

Guardrail: Tell candidates when an interview is recorded or processed by AI, keep humans responsible for hiring decisions, and audit whether pass rates or correction rates differ materially between candidate groups.

Hiring and Interview Insight

The strongest case for AI interviewing may be consistency, not replacement

A new natural field experiment randomly assigned 70,000 applicants to interviews run either by human recruiters or AI voice agents. In both groups, human recruiters reviewed the information and made the hiring decisions.

Applicants in the AI-interview group were 12% more likely to receive offers. The improvement also carried into higher job starts and retention, with no reported decline in the productivity of hired workers. Transcript analysis suggested that the AI interviews were more structured and consistent while still responding to individual answers, allowing them to collect more hiring-relevant information. It is a working paper, so the result should be treated as strong new evidence rather than the final word on every role or population.

The important distinction is that the AI collected information, while people retained decision authority.

One change to test this week:

Create a structured first screen for one repeat-hire role:

• Five core questions asked of every candidate

• Two permitted follow-ups for each question

• Behavioural anchors for scores one, three, and five

• AI-generated transcript and draft evidence summary

• Mandatory human approval before progression or rejection

• A candidate feedback question immediately afterwards

Measure: completion rate, information coverage, human override rate, offer rate, candidate satisfaction, subgroup pass rates, early retention, and hiring-manager satisfaction.

If the process gets faster but candidates hate it, or if output varies unfairly between groups, it is not a successful pilot. Efficiency is one metric, not the whole scorecard.

Funding Watch

All announced during the last seven days. Treat these as hiring and sourcing signals, not proof that every company will immediately open dozens of roles.

River AI | $1.1 billion

Founded by xAI co-founder Igor Babuschkin, River AI is building tools that let enterprises customise open-weight models on their own data. It says reinforcement-learning runs can be completed in 15 to 20 minutes without a dedicated infrastructure team and at two to four times lower cost than closed alternatives. Likely hiring: reinforcement learning, distributed training, enterprise data, inference infrastructure, and applied AI.

Neros Technologies | $250 million Series C | $2.5 billion valuation

The defence technology company is scaling autonomous and interceptor drones, including its Archer AI and Bandit programmes, plus the control stack for multiple drones. Likely hiring: autonomy, perception, embedded systems, RF, manufacturing, flight test, and defence deployment.

Silicon Data | $30.5 million Series A

Silicon Data provides independent GPU pricing and performance benchmarks, including data intended to support CME Group's planned GPU futures market, subject to regulatory approval. Likely hiring: market data, distributed systems, GPU benchmarking, quantitative engineering, data products, and infrastructure economics.

Mindgard | $30 million Series A

The UK-founded AI security company says it has uncovered more than 150 high-impact security and safety vulnerabilities across widely used AI products. The funding will support product, engineering, sales, marketing, and international expansion. Likely hiring: AI red teaming, application security, runtime protection, security research, and enterprise go-to-market.

Cytix | $7 million Series A

The Manchester cybersecurity company is building a change-risk platform for the higher volume of software updates created by AI-assisted development. It is targeting enterprise and regulated customers and is available through partnerships with NCC Group and KPMG. Likely hiring: security engineering, developer tooling, risk analytics, compliance integrations, and enterprise sales.

Quick Bytes

• The AI world may split into formal blocs. A draft US letter reportedly tells 35 partner countries that membership of a US-led AI coalition would be incompatible with joining Beijing's competing framework. Supply-chain, export-control, and regional policy expertise will become more important to AI infrastructure decisions.

• Open cyber models are approaching restricted systems in some tests. Z.ai said GLM-5.3 scored 84.5% on CyberGym versus 83.8% for Anthropic's Mythos 5, although the claims have not been independently verified and GLM-5.3 remained well behind on exploit development. Z.ai plans a safety review and trusted-access controls before wider release.

• Autonomous trucking opens another major market. Aurora and Kodiak received permits to test heavy self-driving trucks on California public roads with human safety operators. Expect continued demand for autonomy, simulation, safety systems, mapping, fleet operations, and regulatory engineering.

• Crypto security gets another reminder. SafePal disclosed unauthorised access affecting order information for about 39,798 customers, including names, addresses, and purchase data.

What to do this week

1. Add AI economics to one senior technical interview

Give candidates a workload with different model prices, latency, and quality. Ask them to propose a routing policy and define cost per successful task.

Metric: quality-adjusted cost, not token cost.

2. Audit every internally created AI or vibe-coded application

Identify its owner, production users, secrets, dependencies, data access, monitoring, security review, and rollback path.

Metric: percentage of live applications with a named owner and documented controls.

3. Run the structured-interview pilot

Use the Screenloop test above, with humans reviewing every AI-drafted scorecard.

Metric: time to feedback, correction rate, evidence coverage, and candidate experience.

4. Build a five-company sourcing watchlist

Map River AI, Neros, Silicon Data, Mindgard, and Cytix before hiring announcements become crowded.

Metric: 25 relevant profiles and five warm conversations across the most likely hiring functions.

This week's signal is not simply that AI has more money. It is that the industry is becoming more financially engineered, more regionally fragmented, and more demanding about production economics.

The winners in the next phase will not just build impressive demos. They will know what the system costs, what can go wrong, which market rules apply, and how to keep the whole thing useful when the pilot becomes a product.

That is all for this week's Tech Talent Drop. Stay informed, and see you next week.